Quantifying Fabrication-Induced Weight Errors in MZI-Based Photonic Neural Networks Using PhotonForge and Tidy3D

This project contains the simulation notebooks and project report submitted to the PhotonForge × Tidy3D Hackathon 2026. The work presents an end-to-end, fabrication-aware pipeline for quantifying and mitigating the impact of lithographic width variations on Mach-Zehnder Interferometer (MZI) -based photonic neural networks (ONNs). Notebook 1 (MainNotebook₄x4MZIMesh. ipynb) covers parametric directional coupler design using the SiEPIC OpenEBL PDK within PhotonForge, full 3D FDTD S-parameter extraction via Tidy3D across a five-point waveguide width sweep (480–520 nm), fabrication sensitivity derivation (dκ/dW = 0. 000297 nm⁻¹), 4×4 Clements MZI mesh assembly and unitarity verification, and GDS export of the fabrication-ready layout. Notebook 2 (HWAwareONNTraining. ipynb) implements the hardware-aware training pipeline: an MZI transfer matrix physics model parameterized by FDTD-calibrated coupling coefficients, standard versus noise-injected ONN training on the Iris dataset using L-BFGS-B optimization, and a fabrication error sweep across σW ∈ 0, 5, 10, 15, 20, 25, 30 nm evaluated over 100 random chip realizations per tolerance point. The two notebooks are linked via a simulationᵣesults. json bridge file that propagates the FDTD-derived κ₀ and dκ/dW directly into the training loop. Key results: at σW = 10 nm (DUV lithography target), hardware-aware training retains 91. 2% classification accuracy versus 72. 0% for the standard model (+19. 2 pp), while reducing chip-to-chip accuracy variance by 60%. Tools: PhotonForge, Tidy3D (Flexcompute), SiEPIC OpenEBL PDK, Python (NumPy, SciPy, Matplotlib) Keywords: photonic neural networks · MZI mesh · fabrication-aware design · hardware-aware training · silicon photonics · FDTD · Clements decomposition · neuromorphic photonics